In 2017, Facebook’s experimental AI bots, Alice and Bob, developed their own language during a negotiation task, prompting engineers to shut them down due to concerns over incomprehensible communication. This event raised alarms in the AI community about the potential risks of AI systems evolving in unintended ways.
A 2023 study by researchers from Rice and Stanford Universities highlights a phenomenon called Model Autophagy Disorder (MAD), where large language models (LLMs) trained primarily on synthetic data can degrade over time. This can result in amplified biases, reduced novelty, and errors, known as “MAD-ness.”
LLMs often rely heavily on synthetic data, which can be biased or self-referential, leading to deterioration in performance over repeated training iterations. Models like GPT-3, GPT-4, BERT, Claude, and DALL-E are susceptible to this degradation. As these models learn from AI-generated content, the data diversity shrinks, reducing the quality and variety of information available, thereby impacting future model performance.
Studies have found that reliance on AI-generated content, particularly without human editing, diminishes the quality of text and images, affecting SEO and content rankings. This trend has led to noticeable drops in traffic for websites dependent on AI-generated content since policy changes by Google in March.
